AI Agents for MaritimeShipping Operations 2026

If you run ships, 2026 is a strange year. The regulatory item that dominates every conference agenda has not happened, may not happen in December, and is being sold to you as a deadline. The instrument that genuinely changed in 2026 — the MASS Code — was adopted in May and took effect in July with a fraction of the coverage. And the compliance obligation that is actually costing money right now, the EU Emissions Trading System, is routinely described with the wrong year attached to it.

This article is written for a fleet operations or technical director who has to make budget decisions against that picture. It starts with the regulation because in maritime the regulation determines which AI work is worth doing. Then it gets to the part that matters commercially: what AI actually delivers on a vessel and in a shore office, which of the published savings numbers survive scrutiny, and which ones evaporate as soon as you ask a single follow-up question.

Here is the calendar as it stands. Legal status is stated separately from press coverage on purpose, because the gap between the two is where most bad procurement decisions in this sector originate.

Maritime regulatory calendar as of August 2026. Net-Zero Framework mechanics and FuelEU cycle dates are trade-sourced; MASS Code, EU ETS scope and MSC decisions are from primary sources.

Get this right and you will be more accurate than most of the trade press. The IMO Net-Zero Framework was approved at MEPC 83, held 7 to 11 April 2025. Approval is not adoption. Adoption was scheduled for an extraordinary session, and at that session — MEPC/ES.2, 13 to 17 October 2025 — the decision was adjourned for twelve months on a vote of 57 in favour of adjournment to 49 against.

MEPC 84 met in London from 27 April to 1 May 2026. It preserved the framework as the basis for negotiation, but produced no compromise between the camp arguing for adoption with minimal change and the camp arguing that the market is not ready. The adjourned session now resumes for a single day on Friday 4 December 2026, immediately after MEPC 85, which sits from 30 November to 3 December.

The timing arithmetic is the part that changes your capital planning. Entry into force follows sixteen months after adoption under the tacit acceptance procedure. DNV's frequently quoted line — 1 March 2028 at the earliest — was calculated from an assumed October 2026 adoption. Run the same sixteen months from a December 2026 adoption and you land at roughly April 2028. March 2028 is a floor, not a date, and it is a floor that only exists if the December vote actually carries.

The mechanics below are reported by trade and class sources rather than confirmed as law, because they describe an instrument that has not been adopted. They are worth knowing because they define the shape of the exposure, not because they define an obligation.

Reported Net-Zero Framework mechanics. Every figure in the middle column describes an instrument that is neither adopted nor in force.

One detail deserves more attention than it gets: both tier prices are reported as fixed for 2028 to 2030 only. Every retrofit and alternative-fuel model we have reviewed that uses these numbers extrapolates USD 380 flat through 2035, and none of them label that as an assumption. It is not conservatism, it is an unexamined input. If you are going to build the case, build it with 2031 onward as an explicit unknown and show the sensitivity to whoever signs the capital request.

On the politics: the US administration opposed the framework and, during the October 2025 session, threatened sanctions, visa restrictions, tariffs and port fees against states supporting it. Reporting from MEPC 84 describes a less vocal posture. Both of those are trade-sourced characterisations rather than measurable facts, and neither should carry weight in a financial model. What should carry weight is that a 57 to 49 vote is not a stable majority in either direction.

Build the data capability, not the compliance product. The measurement work — reconciled fuel consumption, verified well-to-wake intensity, auditable voyage records — is required by EU ETS and FuelEU today and would be required by the Net-Zero Framework if it is ever adopted. That work has a positive return under every branch of the December vote. A product built specifically for an unadopted framework has a return under exactly one.

— Frenchy Digital planning principle

While the sector argued about the Net-Zero Framework, the IMO adopted the MASS Code on 22 May 2026 at MSC 111. It took effect on 1 July 2026. It is the first international instrument covering Maritime Autonomous Surface Ships, and it received a fraction of the attention the unadopted framework received.

The scope is deliberately narrow and worth stating exactly. The Code is non-mandatory. It applies to cargo ships under SOLAS Chapter 1 only — passenger ships and fishing vessels are excluded. It covers four degrees of autonomy. It requires flag-certified Remote Operations Centres. And it requires a designated master, ashore or aboard, with a genuine ability to intervene.

The timeline for the mandatory version is long. A mandatory Code is targeted for adoption by 1 July 2030 and entry into force on 1 January 2032. MSC 112 in December 2026 is expected to set the framework for the Experience-Building Phase between now and then. In practical terms, 2026 to 2030 is a supervised-trial period conducted under flag agreement, not an era of commercial autonomy.

Four substantial questions remain unresolved, and each is a commercial risk rather than a technical one:

• Liability allocation: Where responsibility sits between owner, operator, Remote Operations Centre and flag has not been settled. Until it is, an autonomous or remotely supervised operation is a bespoke contractual negotiation for every deployment.
• Software-developer and ROC fault: Whether and how a software supplier or a remote operations centre bears fault after a casualty is open. This is the single largest unpriced exposure in the autonomy value chain, and it is not one your P&I cover was written for.
• Evidence collection: What must be recorded, in what form, and for how long, to reconstruct a casualty involving an autonomous system. If you are running trials, over-record now — logging you cannot go back and add is the cheapest insurance in this space.
• Search and rescue: SAR obligations for and by MASS remain unresolved, which matters for any operation in a trade lane where you would ordinarily be a first responder.

This is where the money is in 2026, and it is also where imprecision is most expensive. The EU Emissions Trading System phase-in for shipping is described by the Commission in terms of the surrender year. Trade press frequently describes it in terms of the emissions year. Both are defensible. They are not interchangeable, and mixing them produces a document that reads as an error to anyone who knows the rule.

EU ETS maritime surrender phase-in, stated by surrender year. Source: European Commission.

So: during 2026 you surrender 70 percent of your verified 2025 emissions. The statement that the obligation is 100 percent from 1 January 2026 refers to the 2026 emissions year, which is surrendered during 2027. If your board paper says 100 percent for 2026 without that qualification, someone in the room will assume you have the year wrong.

Scope, which is stable and worth restating: ships above 5,000 GT regardless of flag, covering 100 percent of intra-EU voyages and time at berth, and 50 percent of voyages beginning or ending outside the EU.

FuelEU Maritime, Regulation (EU) 2023/1805, has applied since 1 January 2025 to the same 5,000 GT population. The greenhouse gas intensity target is minus 2 percent against a baseline of 91.16 gCO2e/MJ, tightening to minus 6 percent in 2030 and minus 80 percent in 2050. The penalty is EUR 2,400 per tonne of VLSFO-equivalent energy in deficit. Trade reporting puts the first compliance cycle at reports due 31 January 2026 with verification by 31 March 2026.

Note that FuelEU and EU ETS are separate obligations landing on the same vessels with different baselines, different units and different reporting mechanics — 91.16 gCO2e/MJ for FuelEU against a reported 93.3 gCO2eq/MJ baseline in the unadopted IMO framework. Any system that stores a single emissions number per voyage will fail all three regimes at once. The data model has to carry the underlying consumption and fuel-pathway detail, and derive each regime's figure from it.

One live item to watch rather than to plan around: the Commission tabled COM(2026) 616 final on 17 July 2026, a revision of the ETS Directive and the MRV Maritime Regulation. We have verified that the document exists and that this is its date. We have not verified its contents, and neither should you until you or your counsel have read it. Reporting suggests it addresses an IMO deduction mechanism and a revenue-recycling mechanism; that is worth knowing and not worth budgeting against.

This is the section where the article earns its keep, because voyage optimisation is where most maritime AI budget goes and where the published evidence is at its weakest.

Start with the one genuinely independent anchor. The IMO-Norway GreenVoyage2050 Low Carbon Global Industry Alliance study, executed by MarineTraffic and EERA and published in June 2022, analysed 339,390 containership voyages from 2019 AIS data covering 43.97 million tonnes of fuel. It found a 14.16 percent mean per-voyage fuel saving when optimising across the whole voyage.

The two secondary results are more useful than the headline. Optimising only over the last 24 hours yielded 5.90 percent. Over the last 12 hours, 4.23 percent. The value is in the whole voyage. Any product that only intervenes on the approach is fishing in a much smaller pond, and you should ask precisely where in the voyage a vendor's system actually acts before you accept its percentage.

Now the vendor landscape, sorted by how much weight each claim can carry.

Published voyage-optimisation savings claims and their provenance, August 2026. [VENDOR] marks a figure published by the seller.

A second Blue Visby trial with Marubeni in November 2024 covered 16 LPG tankers over 40 voyages and reported roughly 29 percent average with speeds falling from 13.6 to 10.6 knots — but those figures are described as projected or estimated from models rather than measured. Bureau Veritas issued a validation on 22 May 2025, and the scope of that validation is the detail that matters: it validated the methodology for estimating effects, explicitly not measured savings. A class society validating an estimation methodology is a genuine and useful thing. It is not verification that the fuel was saved.

The failure of this genre is not dishonesty, it is unstated counterfactuals. Every savings percentage answers the question “compared to what would otherwise have happened?” and nobody knows what would otherwise have happened. Service speed, contractual speed, historical fleet average and last year's same voyage are all defensible baselines that produce materially different numbers.

Ask one question of every fuel-savings claim: against which baseline, and who measured it? A vendor who answers precisely has done the work. A vendor who cannot answer has given you a number that cannot be audited — which means it also cannot be defended to your charterers, your lenders or your board.

— Frenchy Digital procurement principle

The practical procurement implication is straightforward. Do not buy on a percentage. Buy on a measurement design: agree the baseline in writing before the trial, agree who computes the result, agree the vessel set and the period, and agree what happens if the result comes in below the vendor's published range. Vendors who have real technology accept that. Vendors who do not, do not.

The IMO and Low Carbon GIA Just In Time Arrival Guide contains the most useful set of framing numbers in maritime decarbonisation, and they are worth reading with attention to which are measured and which are modelled.

Just-in-time arrival evidence from the IMO, GloMEEP and Low Carbon GIA work. Note which rows are simulation and which are observation.

Read those rows together and the picture is consistent. Roughly 15 percent of world-fleet marine fuel is burned at port stays, at anchorage, or below one knot. Ships wait at anchorage up to 9 percent of their time. And yet approach-only optimisation at a single large port models out at 4 percent of last-12-hour emissions. The waste is enormous and the fix is not primarily technical.

The reason is contractual and organisational. A requested time of arrival is only actionable if the terminal will commit to a berth window, if the port agent, pilots, tugs and linesmen are aligned to it, and if the charter party does not oblige the vessel to proceed with utmost dispatch regardless. Slowing down to arrive on time can conflict with the charterer's instructions and with demurrage incentives that reward arriving early and waiting. Nobody in that chain is paid to solve the whole problem.

Set expectations accordingly. An agent that drafts and chases port-call correspondence does not deliver 14 percent fuel savings. It delivers a higher rate of confirmed berth windows earlier in the voyage, which is the precondition for whole-voyage optimisation to have anything to optimise against. Measure it on confirmation lead time and on the share of voyages with a firm window more than 24 hours out — not on fuel, which has too many confounders to attribute cleanly.

If you want a clean case study in why digital transformation in shipping runs slower than the announcements suggest, it is the electronic bill of lading. The commitment was specific, public, backed by the largest carriers in the world, and supported by a credible savings estimate. Adoption still moved at a fraction of the promised rate.

eBL commitments versus measured adoption. Sources: DCSA carrier commitment, DCSA interoperability annex, and independent legal analysis published September 2025.

The numbers are worth restating plainly. Nine carriers committed on 15 February 2023 to 50 percent of original bills of lading electronic within five years and 100 percent by 2030. DCSA estimated USD 6.5 billion in direct savings and USD 30 to 40 billion in annual trade growth. As of August 2025, roughly 11 percent of bills of lading were issued electronically — up from around 1 percent in 2021, which is genuine progress, and a long way from the trajectory the commitment implied.

DCSA's own diagnosis is the interesting part. It attributes the gap to behavioural soft barriers, with nearly one third of stakeholders waiting on others to move first. Not technology. Not law — the UK Electronic Trade Documents Act 2023 gave electronic documents parity under English law, and the Netherlands joined France and Germany by 2025. Not even platform lock-in any more: on 4 June 2026, CargoX, edoxOnline, TradeGo, WaveBL and eTEU adopted DCSA's interoperability annex, ending the requirement that both parties use the same platform.

Practically, that means the highest-return document work is internal reconciliation rather than external transformation. Comparing what the bill says against what the booking said, against what the manifest says, against what the statement of facts records — that is entirely inside your boundary, it is where errors actually cost money, and it does not require a single counterparty to agree to anything.

This section exists because the counterintuitive finding is more useful than the conventional one. Almost every maritime cyber briefing you will read implies that mandatory IMO rules are imminent. They are not, and the evidence points firmly the other way.

Resolution MSC.428(98) was adopted on 16 June 2017. The IMO's own page describes it in recommendatory language. It bites in practice not because it is mandatory but because cyber risk has been expected to feature in safety management systems, verified through ISM Code Document of Compliance verification, since 1 January 2021. Supporting guidelines were reissued as MSC-FAL.1/Circ.3/Rev.3 on 4 April 2025.

What does bind, and only for part of the fleet, is class. IACS Unified Requirements E26 and E27 apply to ships contracted for construction on or after 1 July 2024. E26 covers ship-level cyber resilience across identify, protect, detect, respond and recover. E27 covers supplier-side hardening of computer-based systems. Existing tonnage is outside their scope, which means the honest position for most fleets is that cyber obligations remain a safety-management-system matter rather than a certification matter.

The incident record is where careless writing does the most damage, because the widely repeated figures are frequently the outer edge of a range or an unconfirmed vendor claim. Here is what is actually established.

Maritime cyber incident record, with the commonly miscited element identified for each.

One more figure you will encounter: the claim that maritime cyber incidents rose 103 percent in 2025. That is a security-vendor claim from a published white paper, with no absolute incident counts and no disclosed methodology. It may well be directionally right. It is not something to put in a board paper, and a percentage increase with no denominator should not survive a competent review.

On GNSS and AIS interference, the honest position is that the problem is officially acknowledged and not officially quantified. On 26 January 2026, fourteen European coastal states issued a joint open letter stating that GNSS interference originating from Russia is degrading the safety of international shipping. The letter carries no quantified figures. The frequently cited Hormuz and Gulf numbers — more than 1,100 vessels affected in a single 24-hour period, around 21 new jamming clusters, reported 1 March 2026 — are derived from a commercial maritime-analytics vendor. No independent or official quantification exists. Plan for degraded position data as an operating condition; do not build a business case on a vendor's incident count.

Having spent five sections on what the evidence does not support, here is what it does. The defensible AI work in maritime operations in 2026 is administrative and documentation automation under human review. It is unglamorous, it is where the shore-side hours actually go, and it is the only category where you can build a business case from your own baseline rather than from someone else's percentage.

Agent workloads for a fleet operations and technical department, with the human control point for each — Frenchy Digital, 2026.

Two of those rows deserve honest qualification. Certificate and survey expiry surveillance is mostly a database problem wearing an AI costume. If your certificates are already in a structured register, a scheduled query does the job and you should not pay for a model. The AI value appears only where certificates arrive as scanned PDFs from a dozen counterparties in a dozen formats — which, admittedly, is most fleets.

Technical alarm and condition triage is the weakest item on the list, and we have deliberately kept it out of the table. Vessel predictive maintenance has a thin published evidence base — see the limitations section below. If a vendor is selling you machinery failure prediction with a payback guarantee, ask what independent evidence supports the model, and expect the conversation to get vague.

Every one of the workflows above is well within the capability of current models. None of them is within easy reach of most fleets, because the constraint is not the model — it is getting data out of, and results back into, systems that were not built to be integrated with.

Integration constraints for maritime AI deployments and the architectural response to each.

The write path is the item to test first. Reading from a voyage-management or planned-maintenance system is usually negotiable, sometimes through an API and often through a scheduled export. Writing back frequently is not — either the vendor does not expose it, or the licence terms discourage it, or the data model will not accept a partial update. Before committing to a multi-workflow programme, run a two-week spike that attempts one real write into each target system. It is the cheapest way to discover that the second half of your roadmap is not buildable as specified.

Data quality is the second constraint and it is under-discussed. Noon reports are human-entered and self-reported, and they routinely disagree with flowmeter data and with MRV-verified figures. That disagreement is not a bug to be smoothed away by the agent. It is information, and surfacing it is one of the more valuable things an agent can do. What an agent must never do is silently pick a winner between conflicting sources and present the result as fact.

The corresponding control framework worth mapping against is the NIST AI Risk Management Framework. It is not maritime-specific and it will not satisfy a class society on its own, but it gives you a vocabulary for governance that your charterers' and lenders' security teams already recognise.

Send these before the demo. The answers separate vendors who have measured something from vendors who have marketed something, and the difference is usually visible within two questions.

Frenchy Digital due-diligence question set for maritime AI procurement, 2026.

The pricing question is worth pressing harder in this sector than in others, because the norm is genuinely opaque. No vessel software vendor publishes pricing.ZeroNorth's published optimisation product terms contain no fees at all, and nothing is published for Nautilus Labs, DeepSea, Bearing.ai or Wärtsilä's fleet optimisation solution. That is the market convention rather than evidence of bad faith — but it does mean your only leverage is a written quote against a defined scope, and you should not proceed to a pilot without one.

Finally, ask for one artefact instead of a document pack: a redacted export of the agent's audit log for a single workflow over a single day, showing tool calls, model version and the human decision on each item. A vendor who can produce that has built the logging. A vendor who cannot has not, whatever the security questionnaire says.

Each of these is checkable in under a minute, and each of them has appeared in material we have reviewed for clients this year.

The Frenchy Digital red-flag list for maritime AI buyers, 2026.

A vendor who will not put the baseline, the measuring party, the price and the integration scope into writing before you sign will not put them into the product after you sign either. In a sector where nobody publishes pricing and everybody publishes percentages, the willingness to be specific is the strongest signal you have.

— Frenchy Digital buyer's principle

These are the bands Frenchy Digital uses to scope maritime and fleet operations engagements in 2026. They assume the integration work is in scope rather than discovered later, because discovering it later is what turns a nine-week project into a nine-month one.

Frenchy Digital cost bands for maritime and fleet AI engagements, 2026.

Senior-led delivery runs $150 to $225 per hour, and ongoing retainers run $2,500 to $9,500 per month covering model and dependency upgrades, evaluation set expansion, incident response and a quarterly technical review. Every engagement carries a 30-day post-launch warranty, and you receive a written scope with a fixed-price phased proposal within 5 business days of the discovery call.

The budgeting note that matters most in this sector: the integration substrate is largely a fixed cost, paid once. The first agent pays for the data-flow mapping, the identity model, the audit pipeline and the review queue. The third agent inherits all of it. Fleets that sequence their automation get materially better economics than fleets that run four disconnected vendor pilots in parallel — and in an industry where four vendors will each want their own onboard data feed, that difference compounds fast.

What follows is a list of things we could not substantiate. It is deliberately specific, because a vague admission that the evidence is mixed is not useful to anyone building a budget.

• Vessel predictive maintenance evidence is thin: DNV markets a condition-based maintenance survey arrangement, but its own page names no notation, no adoption count and no quantified benefit. Its digital-twin material names only hull monitoring. We could not find a published, quantified vessel predictive-maintenance result from a class society or an operator. That does not mean the technology does not work; it means nobody has published evidence you can build a case on.
• No AI-specific class guidance from ABS or Lloyd's Register could be verified: DNV-RP-0671, Assurance of AI-enabled systems, published September 2023, is cross-industry recommended practice rather than a class notation. DNV and the University of York launched a four-year AI assurance research project in September 2025 — which is itself evidence that the assurance framework does not yet exist. If a vendor implies class approval for an AI system, ask for the document number.
• No vessel software vendor publishes pricing: Not one. ZeroNorth's published optimisation product terms contain no fees; nothing is published for Nautilus Labs, DeepSea, Bearing.ai or Wärtsilä's fleet optimisation solution. Budget planning in this category is therefore quote-driven, and comparison shopping requires you to run a structured RFP rather than read a price page.
• Crew scheduling AI has no usable evidence at all: We looked and found nothing worth citing — no measured deployment, no published outcome, no independent assessment. We are not going to fill that gap with plausible-sounding claims. If crew scheduling is your priority, treat any vendor claim in that area as entirely unverified and design your own measurement before you commit.
• The most-cited port automation productivity evidence is old and unreplicated: McKinsey's much-quoted finding that operating expenses at automated ports fall only 15 to 35 percent while productivity falls 7 to 15 percent comes from a 2018 publication based on an opinion survey of 40-plus practitioners conducted in 2017. It is roughly eight years old with no replication found. It remains the most honest counterweight to automation marketing, and it is not a current measurement.
• No independent dataset supports automation raising terminal productivity: The World Bank and S&P Global Container Port Performance Index measures vessel time in port rather than moves per hour and makes no automated-versus-conventional comparison. The OECD/ITF's 2021 review found automated terminals at roughly 4 percent of global capacity and cited peer-reviewed work concluding that automation alone cannot be considered to have a highly significant impact on terminal performance. Anyone citing CPPI as evidence either way is overreaching.
• Vendor and independent measurements diverge in a documented way: The cleanest published example: a crane automation supplier advertised a 45 to 55 percent reduction in labour time per crane; independent academic measurement in 2017 found 33 percent. That is the correction factor to keep in mind when you read any unaudited vendor percentage in this sector.
• Non-determinism defeats ordinary change control: A model behind a floating alias can change behaviour without a deployment. Pin model versions, keep prompts and tool definitions in the repository under review, maintain a golden set of real port calls and real emissions reconciliations, run it in CI on every change, and canary upgrades while watching override rates. Otherwise you cannot answer what the system did last quarter.
• Prompt injection is unsolved: Agents reading external correspondence are exposed, and the honest framing is blast-radius reduction rather than prevention. Keep the agent read-heavy, keep external actions behind a human, and assume that an attacker who can send your operations inbox a PDF has partial influence over what the model reads.

None of this argues against building. It argues for building the measurement alongside the agent, picking one workflow where you already own the baseline, and being straight with your own organisation about which numbers are yours and which are someone else's marketing. Operators who instrument the before-state get value from AI. Operators who buy a percentage do not.

And the boundary holds throughout: these are administrative and documentation systems operating under human review. Speed instructions, demurrage positions, emissions submissions, safety classifications and anything touching navigation stay with a named person. The MASS Code makes the same argument from the other direction — even in an instrument written specifically for autonomous ships, the designated master must be able to intervene.